Humanoid Policy Viewer

A browser-based viewer and test bench for humanoid locomotion policies. It runs a MuJoCo WebAssembly simulation of the Asimov robot and drives it with an ONNX policy through onnxruntime-web, so a trained policy can be tried without a GPU or a training stack.

• Try a policy: load one from Hugging Face with a single command, from a local policy library, or use the bundled example.
• Drive and stress it: set velocity commands, push the robot, and change friction, armature, gains, gravity and other physical parameters live.
• Score it: npm run benchmark <policy> runs a benchmark (locomotion, pushes, friction, long walks) from a local folder or a Hugging Face dataset, on your machine. Build your own benchmark in the app's Tests & Benchmark panel: tests, repeats, seeds and leaderboard tasks.
• Reads from Hugging Face, never writes: policies and benchmarks are downloaded; publishing a benchmark and submitting results to the Hub leaderboards are done with cyclotron, which uses this viewer as its benchmark harness.
• Faithful setup: the robot is the canonical asimov-1 model, and each policy's gains, action scale and torque limits are read from its training env.yaml, so what you see matches how it was trained.

To train your own policy, use cyclotron. See Supported policies for the inputs and outputs the viewer expects.

Quick start

Run a policy from Hugging Face use the format npm run hf <policy>

For Example:

npm run hf Menlo/asimov1-locomotion-0818

<policy> is the Hugging Face repo id (org/name) or its URL. This works straight from a fresh clone: it installs dependencies, fetches the Asimov robot model, downloads the policy and opens the viewer with it selected. It needs Node, git and bash. See Run a policy from Hugging Face.

<policy> can also be a folder or .onnx file on disk holding the same files, for example npm run hf ./my-policy; see Run a policy from a folder. Give several policies, or a folder of them (such as a training run's logs), to list them all in the viewer's policy dropdown: npm run hf ./logs Menlo/asimov1-locomotion-0818. A benchmark can be given in any order beside them: npm run hf <policy> datasets/<org>/<name> opens a Hub benchmark (read-only), npm run hf <policy> ./my-benchmark a local one.

Run local policies

To run policies you have on disk, or to work on the viewer itself:

npm run dev                  # start the viewer on the bundled example policy
npm run dev ./logs           # ... also listing every policy folder below ./logs

Like npm run hf and npm run benchmark (which also installs Chromium), it installs the dependencies and fetches the Asimov robot model on the first run. Then open the printed URL. It opens the example benchmark benchmarks/default/ in the Tests & Benchmark panel; give benchmark=<dir> for another. To add your own policy, put it in a folder under policies/ (git-ignored, and not built into dist/):

policies/my-policy/
├── policy.onnx
├── env.yaml       the training run's gains, action scale and torque limits
└── agent.yaml

It appears in the policy dropdown the next time the page loads. To serve policies from somewhere else, see Policy library.

npm run dev listens on all network interfaces, so other machines on your network can reach it, and does not open a browser. npm run hf listens on localhost only and opens the browser. Both take port=<n>.

Where policies come from

Every policy the viewer can run appears in its policy dropdown. They come from these sources:

Source	Where the files are	How you get it
Bundled example	public/examples/checkpoints/asimov/model_aug_18_1/, in this repo	The default with npm run dev
Hugging Face	~/.cache/humanoid-policy-viewer/hf/<org>__<name>/, outside the repo	npm run hf <policy>, or npm run dev <policy>; npm run dev lists every one already downloaded
Local folder	policies/<name>/ in this repo (git-ignored), or wherever HPV_POLICY_LIBRARY_DIR and HPV_POLICY_ROOTS point	Put the files there, then npm run dev
Any folder on disk	Wherever it is; linked, not copied	npm run hf <folder> or npm run dev <folder>; a folder of policies lists every policy folder below it

In every case the gains, action scale and torque limits come from the env.yaml next to the .onnx; see Policy config. The robot is the same everywhere: the asimov-1 submodule.

Docs

Topic
Run a policy from Hugging Face	npm run hf: options, private repos, where downloads are cached
Policy config	What reference_policy_config.json and a policy's env.yaml each control, and which wins
Policy library	Serving policies from outside public/ (HPV_POLICY_* variables)
Benchmarking	npm run benchmark: running a benchmark, reading results, versions, parallelism
Benchmarks	The benchmark format (tests, benchmark.yaml, metrics), creating and changing one, and the inspect command publishing tools read
Adding a robot	Bringing your own MJCF, policy and motion clips
Architecture	How the app, server and commands fit together; words; every HPV_* environment variable; old → new names
Scenes	The asimov-1 submodule and how actuators are added at load

Project structure

• src/views/Demo.vue - UI controls for the live demo
• src/simulation/main.js - bootstraps MuJoCo, Three.js renderer, policy loop, and metric sampling hook
• src/simulation/mujocoUtils.js - scene/policy loading utilities and filesystem preloading
• src/simulation/policyRunner.js - ONNX inference wrapper and observation pipeline
• public/examples/scenes/ - MJCF files + meshes staged into MuJoCo's MEMFS; the Asimov robot is the asimov-1/ git submodule
• public/examples/checkpoints/ - policy config JSON, bundled ONNX files, and motion clips
• src/views/BenchmarkResults.vue - the results page (charts, per-repeat outcomes, tasks)
• scripts/ - the commands: dev.mjs, hf.mjs, run-benchmark.mjs, and inspect.mjs (a benchmark folder or a policy as JSON, for publishing tools)
• scripts/startViewer.mjs, scripts/viewerServer.mjs - how every command starts the dev server, and the server's routes (/api/..., /policy-library/)
• src/benchmark/ - benchmark protocol, test-row and benchmark.yaml formats, scoring; run.js plans cells and builds result.json for both the CLI and the in-app Run, and resultView.js turns a result.json into what the results page and report.html show
• benchmarks/ - local benchmarks: default/ (tracked; what npm run dev opens and what new benchmarks are copied from), and, gitignored, benchmarks made with the app's New benchmark
• benchmark_results/ - (gitignored) saved benchmark results
• test/fixtures/smoke-benchmark/ - a tiny benchmark for tests and the golden check
• an optional policy library served by local Vite middleware under /policy-library/

License and acknowledgements

This repository is a fork of humanoid-policy-viewer by Qingzhou Lu (Axellwppr), a motion-tracking viewer for the Unitree G1, and still carries that viewer's G1 scene and motion-tracking code. The upstream project's demos: Humanoid Policy Viewer, GentleHumanoid Web Demo.

Both the original project and Menlo Research's additions are licensed under the BSD 3-Clause License:

• Original code, author-created motions, and the author-trained policy weights at public/examples/checkpoints/g1/policy_latest.onnx — copyright © 2026 Qingzhou Lu.
• Menlo Research modifications — changes made after upstream commit 0490320, including Asimov support, the benchmark, and Hugging Face model loading — copyright © 2026 Menlo Research Pte. Ltd.

Third-party code, libraries, robot descriptions, meshes, and third-party motion data are not relicensed by this grant. They remain subject to their respective terms; see Third-party notices for sources and scope. This includes the Asimov robot description in the asimov-1 submodule, which is licensed under CERN-OHL-S-2.0.

This viewer builds on MuJoCo, the MuJoCo WASM community, Three.js, ONNX Runtime, Vue, and Vuetify. We also thank Unitree Robotics and the creators of the motion datasets used in the examples. If you build on this viewer, a link back to this repository is appreciated.